Cryo-Electron Microscopy: Freezing a Molecule Mid-Motion

The Problem Is the Water, Not the Cold

A protein's shape is not a property of the protein alone — it is a negotiation with the water packed around it. Take the water away and the molecule collapses. Freeze it the ordinary way and something worse happens: a growing ice crystal expels everything that is not water, shoving salts, buffer and protein into the shrinking channels between crystals, where they are concentrated, dehydrated and crushed. The ordered ice also diffracts the beam into Bragg spots that bury the specimen.

The way out is not to freeze the water but to vitrify it — to cool so fast the molecules never find their lattice. Viscosity climbs through about 10¹² Pa·s and the liquid simply arrests into a glass: low-density amorphous ice, with no long-range order, holding every molecule in the pose it had in solution.

Water makes this hard. Below its homogeneous nucleation temperature of 235 K, supercooled water crystallises almost on contact; above about 150 K, amorphous ice crystallises from the other direction. The 85 K band between them is water's "no-man's land", where bulk liquid water cannot be kept at all. Vitrification means sprinting across it faster than ice nuclei can grow — then keeping the prize, since the glass softens through its transition near ~136 K and devitrifies to cubic ice as it warms on towards 150 K, so every transfer must hold the grid below 123 K (−150 °C).

Three Microlitres, a Blot, and a Plunge

The support is a 3.05 mm copper mesh carrying a perforated carbon film — a Quantifoil R1.2/1.3 grid has 1.2 µm holes separated by 1.3 µm of carbon. Carbon is hydrophobic, so the grid is glow-discharged in a low-pressure air plasma for 15–60 s to make it wettable. All-gold grids (UltrAuFoil) help where beam-induced motion limits the data: carbon on copper contracts differently from its support on cooling, crinkling the foil and storing stress the beam then releases as movement.

About 3 µL of purified protein goes on. Inside a Vitrobot chamber near 4 °C and close to 100 % humidity, filter paper blots the grid for 3–6 s, leaving a free-standing aqueous film spanning each hole only 20–100 nm thick — thinner than a wavelength of green light. The robot then fires the grid into liquid ethane.

Ethane, not nitrogen, and the reason is thermodynamic. Liquid nitrogen at 77 K (−196 °C) sits exactly at its boiling point, so a warm grid entering it wraps itself in an insulating vapour blanket — Leidenfrost film boiling. Liquid ethane, held at ~90 K (−183 °C) by that same nitrogen, is roughly 95 K below its own boiling point of 185 K: it cannot boil against the sample, so it stays in liquid contact and drinks heat away at ~10⁵ K/s.

The Arithmetic of Outrunning Ice

Work the numbers. A drop of ~190 K at ~10⁵ K/s takes about 2 ms. Only part of that journey is dangerous: the span from 235 K to 150 K is 85 K wide, so the film crosses no-man's land in 85 / 10⁵ ≈ 0.85 ms — and in the thinnest films, where rates approach 10⁶ K/s, nearer 85 µs. Ice nuclei cannot reach critical size and grow through a 50-nm film in that time, so the liquid runs out of clock and glasses over instead.

Thinness is what buys the rate. Water's thermal diffusivity α is about 1.4 × 10⁻⁷ m²/s, so a 100 nm film equilibrates internally in roughly L²/α = (10⁻⁷ m)² / 1.4 × 10⁻⁷ m²/s ≈ 70 ns. There is essentially no temperature gradient inside the film; the whole 190 K drop is limited by heat leaving the grid for the ethane. That is also why plunge-freezing vitrifies at most about ten micrometres of water — a thin film or a single adherent cell, but not a tissue biopsy, which needs high-pressure freezing at ~2,100 bar to suppress nucleation and extend vitrification to a couple of hundred microns.

On a real grid the gradient shows: the film is thinnest at the centre of each hole and thickens towards the rim and the grid bars, where it can crystallise outright. The microscopist picks the good squares and ignores the rest.

300 kV Electrons and a Dose You Can Only Spend Once

Accelerated through 300 kV, an electron has a relativistic de Broglie wavelength of 1.97 pm, about a hundred times smaller than the diameter of an atom, so resolution is never diffraction-limited. It is limited by lens aberrations and, far more severely, by the fact that the specimen is destroyed while you look at it.

Protein and vitreous ice scatter almost identically, so amplitude contrast is only a few percent. Images are deliberately recorded out of focus (typically 0.5–3 µm underfocus) so interference converts the specimen's phase shift into visible intensity. The price is the contrast transfer function, which oscillates, flips sign and zeroes out whole bands of detail, so every micrograph must be CTF-fitted and corrected before averaging.

Damage sets the exposure budget. In light atoms the inelastic events that break bonds outnumber the elastic ones carrying structure by roughly three to one, each depositing ~20 eV into a molecule whose C–C bonds cost 3.6 eV. A total exposure of ~40 e−/Ų is a dose of order 10⁸ Gy — several times the ~3 × 10⁷ Gy Garman limit past which a cryo-cooled protein crystal's diffraction is already unacceptably degraded, and past even the ~4 × 10⁷ Gy at which its summed intensity halves. Past about 10 e−/Ų the finest detail is gone; side chains blur while the envelope survives.

Direct electron detectors (Gatan K3, Thermo Fisher Falcon 4i) fixed this twice over. They register electrons directly in silicon, with a detective quantum efficiency near 0.8–0.9 at low spatial frequency against roughly 0.3 for the scintillator-coupled CCDs they replaced; and they read out fast enough to record the exposure as a movie of about 40 frames at ~1 e−/Ų each. That matters, because beam-induced motion displaces particles by tens of Ångströms mid-exposure, so frames are aligned patch-by-patch and dose-weighted: early frames carry the high-resolution signal, late ones only the low.

From a Million Shadows to One Structure

Each particle in a micrograph is a 2D projection through the whole molecule, at an orientation nobody told you and a signal-to-noise ratio well below 1 — grey fog with hints in it, not a picture of a molecule.

The rescue is the projection-slice theorem: the 2D Fourier transform of a projection is a central slice through the 3D Fourier transform of the object. Assign the right three Euler angles and two shifts to enough projections and Fourier space fills with slices, which invert to a volume. RELION (maximum-likelihood) and cryoSPARC (stochastic gradient descent) do this iteratively over 10⁵–10⁶ particle images, picked by template matching or a neural network and cleaned by 2D classification.

Averaging is what beats the noise, and noise falls as 1/√N: 10⁴ to 10⁶ particles is a tenfold gain in signal-to-noise. Resolution does not improve linearly with N, though — the Rosenthal–Henderson analysis has the required count rising roughly exponentially in 1/d², which is why the last half-Ångström costs more data than the first four.

One detail is routinely misstated: a cryo-EM map is a Coulomb potential map, not electron density. Electrons are scattered by the nucleus screened by its own electron cloud, so the method is directly sensitive to charge, and below about 1.5 Å individual hydrogens appear. And because the million particles are not identical, 3D classification and heterogeneity methods (multi-body refinement, 3DVA, cryoDRGN) sort them into distinct conformational states — the honest sense in which a molecule is caught mid-motion: not a movie, but a frozen census of an ensemble.

Hardware, and How a Resolution Number Gets Certified

The workhorse is a Thermo Fisher Titan Krios: 300 kV, field-emission gun, constant-power lenses, and an autoloader holding a cassette of twelve grids under cryogenic vacuum, so the specimen never sees room temperature again. An energy filter (Gatan BioQuantum, Thermo Fisher Selectris X) with a 10–20 eV slit discards inelastically scattered electrons that contribute only fog. Column vacuum runs near 10⁻⁷ mbar, and the cold specimen acts as a cryopump — which is why frost is the perennial enemy of the cold chain.

Resolution is certified, not asserted. The particle set is split into two halves refined completely independently, and the maps compared shell by shell by Fourier shell correlation; the quoted gold-standard resolution is where FSC falls to 0.143, a threshold set so the combined map correlates with the true structure at about 0.5. Because the halves never exchange information, overfitting cannot flatter the number. Serious depositions add local-resolution maps — the periphery is almost always worse than the core — plus map-to-model FSC and scores such as EMRinger, with maps going to the EMDB, coordinates to the PDB and raw movies to EMPIAR.

How It Got Here, and How It Fails

Richard Henderson and Nigel Unwin showed in 1975 that an unstained biological molecule could survive the beam at all, solving bacteriorhodopsin to 7 Å from glucose-embedded 2D crystals. Jacques Dubochet and Alasdair McDowall showed in 1981 at EMBL Heidelberg that a plunged thin water film really does vitrify — diffuse halos in its diffraction, not rings — and the group imaged viruses in Nature in 1984. Joachim Frank supplied the mathematics for aligning and classifying unaligned single particles, beginning with SPIDER in 1981. Direct detectors arrived in 2012–13, and Werner Kühlbrandt named the consequence the "resolution revolution" in Science in 2014. Dubochet, Frank and Henderson shared the 2017 Nobel Prize in Chemistry; in 2020 two groups pushed apoferritin to 1.22 Å and 1.25 Å. Weeks after the genome sequence appeared that year, the prefusion SARS-CoV-2 spike was solved, and the stabilised conformation it revealed became the vaccine antigen.

The failure modes are specific. Let a grid warm past ~136 K in a clumsy transfer and the arrested liquid devitrifies to cubic ice: crystalline reflections appear in the power spectrum and the data is dead. Thick ice adds inelastic fog no averaging removes. The worst offender is the air–water interface, because a 50 nm film is almost entirely surface. Einstein–Smoluchowski gives t ≈ x²/2D; for a 100 kDa protein with D ≈ 5 × 10⁻¹¹ m²/s crossing 25 nm that is only ~6 µs, so during a 3-second blot each particle strikes an interface something like 10⁵–10⁶ times. Many partly denature there or adsorb in one favoured pose, starving Fourier space in one direction and yielding a smeared, anisotropic map. Remedies: detergents, tilted collection at 30–40°, graphene supports, and blot-free piezo dispensers (Spotiton, Chameleon) that go tip-to-ethane in ~100 ms.

The subtlest failure is model bias: reference-based alignment of pure noise reproduces the reference. Henderson's 2013 demonstration pulled a recognisable Einstein's head out of noise aligned against an Einstein template, and a 2013 HIV-1 trimer map was challenged on exactly those grounds — which is why half-sets are refined in isolation.

And there is a floor: alignment gets hard below ~50 kDa, where 52 kDa streptavidin and 64 kDa haemoglobin mark the practical limit. Even time-resolved cryo-EM, spraying milliseconds before the plunge, yields separate stills, never a movie.

Cryo-EM against the methods it is most often mistaken for

Method	What the sample is	What the map actually is	Typical resolution and limit
Single-particle cryo-EM	Purified molecules in a 20–100 nm vitreous film; no crystal	Coulomb potential, averaged over 10⁵–10⁶ separate copies	1.2–4 Å; alignment gets hard below ~50 kDa
X-ray crystallography	A well-ordered crystal, often the hardest year of the project	Electron density, averaged over ~10¹⁵ molecules locked in a lattice	0.8–3 Å; crystal packing can bias the conformation
Cryo-electron tomography	One unique object — a cell, a virion, an FIB-milled lamella	A 3D volume of that single object, built from a ±60° tilt series	2–4 nm; 3–6 Å after sub-tomogram averaging of repeats
Negative-stain EM	Air-dried at room temperature in uranyl salts	A heavy-metal cast of the outside surface	~20 Å; interior invisible, shapes flattened
Solution NMR	Isotope-labelled protein in ordinary liquid water	Distance and angle restraints fitted to an ensemble of models	Atomic, but practically capped near ~30 kDa